Skip to content

RAG playground

Run the RAG types in your browser and watch every step. Pick one or more chunkers and RAG types, choose a question, and each setup runs side by side on its own flow diagram: the steps light up in the order they ran, with step numbers, and the path taken turns orange. Click any step to see what went in and what came out (the chunks found, with their files and scores, the rewritten query, the evidence check).

The playground needs JavaScript.

How it works

  • The real rag-titan code runs in your browser with Pyodide (Python compiled to WebAssembly). The first visit downloads about 10 MB; later visits use your browser's cache.
  • In the default recorded mode, the model replies and embedding vectors are real, recorded once with gemma3:4b and nomic-embed-text on the three sample policies in examples/sample_docs/, then replayed. Search, reranking, the RAG types, the scores and the diagrams are all computed live.
  • Evaluate all questions runs every chosen setup on the six test questions in examples/questions.json and scores them like rag-titan evaluate: did the chunks given to the model contain the evidence, and was the question answered?

Things to try

  1. Chunking changes the answer. Ask When do I need a note from the doctor? with adaptive on word and on paragraph. With word chunks the evidence check says "no" twice and it gives up after 9 steps; with paragraphs it answers in 5. Click Retrieve in the failed run: the right sentence was found, the model just did not trust it.
  2. A retry that works. Ask What is the maximum hotel cost per night? with adaptive + paragraph: the first evidence check fails, the query is rewritten, and the wider search (top_k 4 → 8) succeeds.
  3. Two files needed. Ask the furniture question with simple: it keeps only two chunks, so the second file may never reach the model.
  4. Compare everything. Tick all chunkers and RAG types and press Evaluate all questions.

Live mode: your Ollama, any question, your own text

Work in progress

Live mode does not work on this website yet: Chrome blocks a public website from reaching the Ollama on your computer (its new local network access protection), and the fix is still being made. It does work when you run the docs on your own machine (mkdocs serve). The recorded runs above work everywhere.

Switch Mode to Live with your Ollama to ask any question, on the sample documents or on text you paste, with the models on your own machine. The page talks to Ollama on your computer; nothing is sent anywhere else.

  1. Ollama must allow this page. Start it with:

    OLLAMA_ORIGINS=https://aion7891.github.io ollama serve

    With the Ollama app on a Mac, run launchctl setenv OLLAMA_ORIGINS https://aion7891.github.io once, then quit and reopen the app.

  2. Have a chat model and an embedding model, for example ollama pull gemma3:4b and ollama pull nomic-embed-text. Cloud models (gpt-oss:20b-cloud, ...) work too after ollama signin, if your Ollama account has credits.

  3. Press Connect, pick the models, and Run.

Your browser may ask whether this page may access devices on your local network: allow it, that is your own Ollama. Live runs take a few seconds per setup, and the evidence is only scored for the listed questions on the sample documents. Evaluate all questions is only offered for recorded runs: live it would ask your model about 50 times.